OCR for Text

---

Overview

Supported devices are CPU and GPU only; NPU is not supported.

OCR for Text is a Metro Analytics use case that detects and reads text in images and video streams using the PaddleOCR PP-OCRv4 pipeline. It composes two models:

• PP-OCRv4 Detection (ch_PP-OCRv4_det) -- a lightweight DBNet-based text detector that locates text regions in the frame.
• PP-OCRv4 Recognition (ch_PP-OCRv4_rec_server) -- the larger "server" CRNN-CTC recognizer variant, which is more accurate than the lightweight mobile variant on stylized or decorative fonts, and converts each cropped text region into a character string.

Both models are converted to OpenVINO IR using the ovc (OpenVINO Model Converter) tool which reads PaddlePaddle models directly. This is the best supported end-to-end OCR stack for OpenVINO.

Typical Metro deployments include:

• Signage Reading -- read platform signs, departure boards, safety notices.
• Document Scanning -- extract text from forms, labels, and ID cards.
• Label Verification -- read package labels or barcodes in logistics.
• Multilingual Support -- PP-OCRv4 supports multiple scripts out of the box.

For license-plate-specific OCR, see the license-plate-recognition use case which includes a specialized plate detector.

---

Prerequisites

Target stack: Intel DLStreamer 2026.3.0 RC1 with the matching OpenVINO GenAI bundle 2026.5.0.0.dev20261030 (OpenVINO Runtime 2026.5.0.dev20261030). Use the native RC1 setup before running these samples; this is a development stack, not a stable release. See the device validation matrix for measured results and device support exclusions.

• Python 3.11+
• Install OpenVINO (latest version)
• Install Intel DLStreamer (latest version)

Create and activate a Python virtual environment before running the scripts:

python3 -m venv .venv --system-site-packages
source .venv/bin/activate

Note: The --system-site-packages flag is required so the virtual environment can access the system-installed OpenVINO and DLStreamer Python packages.

---

Getting Started

Download and Convert Models

Run the provided script to download the PaddleOCR models and convert them to OpenVINO IR:

chmod +x export_and_quantize.sh
./export_and_quantize.sh

The script performs the following steps:

1. Installs dependencies (openvino).
2. Downloads the PP-OCRv4 detection and recognition inference models.
3. Converts both to OpenVINO IR format using ovc.
4. Downloads a sample test image with text, a sample test video (test_video.mp4, a close-up of street name and stop signs), and the PP-OCRv4 character dictionary (ppocr_keys_v1.txt) used to CTC-decode the recognizer's output into text.

Output files:

• ch_PP-OCRv4_det_infer/ -- detection model (OpenVINO IR).
• ch_PP-OCRv4_rec_server_infer/ -- recognition model, server variant (OpenVINO IR).
• ppocr_keys_v1.txt -- character dictionary for the recognizer's CTC decoder.

OpenVINO Sample

The sample below runs the full PP-OCRv4 pipeline across every frame of a video: the detector locates text regions (using an aspect-ratio-preserving resize and a dilation step so a whole word is captured in one box instead of fragments), then the recognizer reads each cropped region and CTC-decodes it into a text string, which is drawn as a solid-background label directly over its box so the highlighted region visibly shows what is written. Change the device string to run on CPU or GPU.

The full sample is in openvino_sample.py; run it with python3 openvino_sample.py or copy the code below.

# SPDX-License-Identifier: MIT
# Copyright (C) Intel Corporation
#
# OCR for Text: OpenVINO Runtime sample.
# See README.md for prerequisites and inputs. Run: python3 openvino_sample.py

import cv2
import numpy as np
import openvino as ov

DET_MODEL = "ch_PP-OCRv4_det_infer/inference.xml"
REC_MODEL = "ch_PP-OCRv4_rec_server_infer/inference.xml"
DICT_FILE = "ppocr_keys_v1.txt"
INPUT_VIDEO = "test_video.mp4"
DET_SIZE = 960

core = ov.Core()

# Change device to "GPU" to run on integrated GPU.
det_compiled = core.compile_model(core.read_model(DET_MODEL), "CPU")
rec_compiled = core.compile_model(core.read_model(REC_MODEL), "CPU")

# CTC label map: index 0 is the blank symbol, followed by every character in
# the dictionary file, followed by a trailing space character.
chars = open(DICT_FILE, encoding="utf-8").read().splitlines()
dict_character = ["blank"] + chars + [" "]


def detect_text_regions(frame, thresh=0.3, pad=4):
    """Return (x, y, w, h) boxes for words/lines of text in a frame.

    Resizing preserves aspect ratio (letterboxed onto a square canvas) so
    text isn't skewed, and dilating the detection map merges nearby
    characters into one box per word instead of one per character.
    """
    h0, w0 = frame.shape[:2]
    scale = DET_SIZE / max(h0, w0)
    resized = cv2.resize(frame, (int(w0 * scale), int(h0 * scale)))
    canvas = np.zeros((DET_SIZE, DET_SIZE, 3), dtype=np.uint8)
    canvas[:resized.shape[0], :resized.shape[1]] = resized

    blob = canvas.astype(np.float32).transpose(2, 0, 1)[np.newaxis] / 255.0
    det_map = det_compiled([blob])[det_compiled.output(0)][0, 0]
    binary = (det_map > thresh).astype(np.uint8) * 255
    dilated = cv2.dilate(binary, np.ones((9, 25), np.uint8))
    contours, _ = cv2.findContours(dilated, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)

    boxes = []
    for c in contours:
        x, y, w, h = cv2.boundingRect(c)
        if w < 10 or h < 5:
            continue
        x0, y0 = max(0, x / scale - pad), max(0, y / scale - pad)
        x1, y1 = min(w0, (x + w) / scale + pad), min(h0, (y + h) / scale + pad)
        boxes.append((int(x0), int(y0), int(x1 - x0), int(y1 - y0)))
    return boxes


def recognize_text(crop, rec_h=48, max_w=320):
    """Resize a cropped text region to the recognizer's input shape and
    CTC-decode the predicted character sequence into a string."""
    h, w = crop.shape[:2]
    if h == 0 or w == 0:
        return "", 0.0
    resized_w = max(1, min(max_w, round(rec_h * w / h)))
    blob = cv2.resize(crop, (resized_w, rec_h)).astype(np.float32) / 255.0
    blob = ((blob - 0.5) / 0.5).transpose(2, 0, 1)[np.newaxis, ...]

    preds = rec_compiled([blob])[rec_compiled.output(0)][0]
    idx = np.argmax(preds, axis=1)
    conf = np.max(preds, axis=1)

    text, scores, prev = [], [], -1
    for i, c in zip(idx, conf):
        if i != 0 and i != prev:
            text.append(dict_character[i])
            scores.append(c)
        prev = i
    confidence = float(np.mean(scores)) if scores else 0.0
    return "".join(text), confidence


def annotate(frame, box, text, confidence):
    """Draw a bounding box and, if any text was recognized, a legible
    label (solid background so it stays readable over any color) above it."""
    x, y, w, h = box
    cv2.rectangle(frame, (x, y), (x + w, y + h), (0, 255, 0), 2)
    if not text:
        return
    label = f"{text} ({confidence:.2f})"
    (tw, th), base = cv2.getTextSize(label, cv2.FONT_HERSHEY_SIMPLEX, 0.7, 2)
    top = max(0, y - th - base - 6)
    cv2.rectangle(frame, (x, top), (x + tw + 6, top + th + base + 6), (0, 255, 0), -1)
    cv2.putText(frame, label, (x + 3, top + th + 2),
                cv2.FONT_HERSHEY_SIMPLEX, 0.7, (0, 0, 0), 2)


cap = cv2.VideoCapture(INPUT_VIDEO)
fps = cap.get(cv2.CAP_PROP_FPS) or 30.0
width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
writer = cv2.VideoWriter(
    "output_openvino.mp4", cv2.VideoWriter_fourcc(*"mp4v"), fps, (width, height))

frame_idx = 0
total_regions = 0

while True:
    ok, frame = cap.read()
    if not ok:
        break
    frame_idx += 1

    for box in detect_text_regions(frame):
        x, y, w, h = box
        text, confidence = recognize_text(frame[y:y + h, x:x + w])
        annotate(frame, box, text, confidence)
        total_regions += 1
        print(f"Frame {frame_idx}: region=({x},{y},{w},{h}) text={text!r} "
              f"confidence={confidence:.2f}", flush=True)

    writer.write(frame)

cap.release()
writer.release()
print(f"Total text regions across all frames: {total_regions}", flush=True)
print("Saved: output_openvino.mp4")

Device targets:

• "CPU" -- default, works on all Intel platforms.
• "GPU" -- Intel integrated or discrete GPU.

Note: Recognition accuracy depends heavily on font, angle, and image quality. Plain block-lettered signage (as in the sample video) decodes reliably; stylized or decorative fonts are harder for a general-purpose OCR model and may not decode perfectly.

Expected Output

DLStreamer Sample

The sample below decodes a video with the DLStreamer/GStreamer stack (decodebin3 ! videoconvert), pulls BGR frames through appsink, runs the PP-OCRv4 text detector on each frame (using an aspect-ratio-preserving resize and a dilation step so a whole word is captured in one box instead of fragments), then runs the PP-OCRv4 recognizer on each cropped region and CTC-decodes the result into text drawn as a solid-background label directly over its box before writing the annotated output to output_dlstreamer.mp4.

Notes on running this sample:

• Export PYTHONPATH so the DLStreamer Python module is importable:

  source /opt/intel/openvino_2026/setupvars.sh
  source /opt/intel/dlstreamer/scripts/setup_dls_env.sh
  export PYTHONPATH=/opt/intel/dlstreamer/python:\
  /opt/intel/dlstreamer/gstreamer/lib/python3/dist-packages:${PYTHONPATH:-}
  

The full sample is in dlstreamer_sample.py; run it with python3 dlstreamer_sample.py or copy the code below.

# SPDX-License-Identifier: MIT
# Copyright (C) Intel Corporation
#
# OCR for Text: DLStreamer pipeline sample.
# See README.md for prerequisites and inputs. Run: python3 dlstreamer_sample.py

import gi

gi.require_version("Gst", "1.0")
from gi.repository import Gst

import numpy as np
import openvino as ov

Gst.init([])

# Import cv2 after Gst.init to avoid GStreamer re-initialization conflicts.
import cv2

INPUT_VIDEO = "test_video.mp4"
DET_MODEL = "ch_PP-OCRv4_det_infer/inference.xml"
REC_MODEL = "ch_PP-OCRv4_rec_server_infer/inference.xml"
DICT_FILE = "ppocr_keys_v1.txt"
DET_SIZE = 960

core = ov.Core()
det_compiled = core.compile_model(core.read_model(DET_MODEL), "CPU")
rec_compiled = core.compile_model(core.read_model(REC_MODEL), "CPU")

# CTC label map: index 0 is the blank symbol, followed by every character in
# the dictionary file, followed by a trailing space character.
chars = open(DICT_FILE, encoding="utf-8").read().splitlines()
dict_character = ["blank"] + chars + [" "]


def detect_text_regions(frame, thresh=0.3, pad=4):
    """Return (x, y, w, h) boxes for words/lines of text in a frame.

    Resizing preserves aspect ratio (letterboxed onto a square canvas) so
    text isn't skewed, and dilating the detection map merges nearby
    characters into one box per word instead of one per character.
    """
    h0, w0 = frame.shape[:2]
    scale = DET_SIZE / max(h0, w0)
    resized = cv2.resize(frame, (int(w0 * scale), int(h0 * scale)))
    canvas = np.zeros((DET_SIZE, DET_SIZE, 3), dtype=np.uint8)
    canvas[:resized.shape[0], :resized.shape[1]] = resized

    blob = canvas.astype(np.float32).transpose(2, 0, 1)[np.newaxis] / 255.0
    det_map = det_compiled([blob])[det_compiled.output(0)][0, 0]
    binary = (det_map > thresh).astype(np.uint8) * 255
    dilated = cv2.dilate(binary, np.ones((9, 25), np.uint8))
    contours, _ = cv2.findContours(dilated, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)

    boxes = []
    for c in contours:
        x, y, w, h = cv2.boundingRect(c)
        if w < 10 or h < 5:
            continue
        x0, y0 = max(0, x / scale - pad), max(0, y / scale - pad)
        x1, y1 = min(w0, (x + w) / scale + pad), min(h0, (y + h) / scale + pad)
        boxes.append((int(x0), int(y0), int(x1 - x0), int(y1 - y0)))
    return boxes


def recognize_text(crop, rec_h=48, max_w=320):
    """Resize a cropped text region to the recognizer's input shape and
    CTC-decode the predicted character sequence into a string."""
    h, w = crop.shape[:2]
    if h == 0 or w == 0:
        return "", 0.0
    resized_w = max(1, min(max_w, round(rec_h * w / h)))
    blob = cv2.resize(crop, (resized_w, rec_h)).astype(np.float32) / 255.0
    blob = ((blob - 0.5) / 0.5).transpose(2, 0, 1)[np.newaxis, ...]

    preds = rec_compiled([blob])[rec_compiled.output(0)][0]
    idx = np.argmax(preds, axis=1)
    conf = np.max(preds, axis=1)

    text, scores, prev = [], [], -1
    for i, c in zip(idx, conf):
        if i != 0 and i != prev:
            text.append(dict_character[i])
            scores.append(c)
        prev = i
    confidence = float(np.mean(scores)) if scores else 0.0
    return "".join(text), confidence


def annotate(frame, box, text, confidence):
    """Draw a bounding box and, if any text was recognized, a legible
    label (solid background so it stays readable over any color) above it."""
    x, y, w, h = box
    cv2.rectangle(frame, (x, y), (x + w, y + h), (0, 255, 0), 2)
    if not text:
        return
    label = f"{text} ({confidence:.2f})"
    (tw, th), base = cv2.getTextSize(label, cv2.FONT_HERSHEY_SIMPLEX, 0.7, 2)
    top = max(0, y - th - base - 6)
    cv2.rectangle(frame, (x, top), (x + tw + 6, top + th + base + 6), (0, 255, 0), -1)
    cv2.putText(frame, label, (x + 3, top + th + 2),
                cv2.FONT_HERSHEY_SIMPLEX, 0.7, (0, 0, 0), 2)


pipeline_str = (
    f"filesrc location={INPUT_VIDEO} ! decodebin3 ! videoconvert ! "
    "video/x-raw,format=BGR ! "
    "appsink name=sink emit-signals=false sync=false"
)
pipeline = Gst.parse_launch(pipeline_str)
sink = pipeline.get_by_name("sink")
pipeline.set_state(Gst.State.PLAYING)

writer = None
frame_idx = 0
total_regions = 0

while True:
    sample = sink.emit("pull-sample")
    if sample is None:
        break
    buf = sample.get_buffer()
    caps = sample.get_caps().get_structure(0)
    width = caps.get_value("width")
    height = caps.get_value("height")

    ok, mapinfo = buf.map(Gst.MapFlags.READ)
    if not ok:
        continue
    frame = np.ndarray((height, width, 3), dtype=np.uint8,
                       buffer=mapinfo.data).copy()
    buf.unmap(mapinfo)
    frame_idx += 1

    for box in detect_text_regions(frame):
        x, y, w, h = box
        text, confidence = recognize_text(frame[y:y + h, x:x + w])
        annotate(frame, box, text, confidence)
        total_regions += 1
        print(f"Frame {frame_idx}: region=({x},{y},{w},{h}) text={text!r} "
              f"confidence={confidence:.2f}", flush=True)

    if writer is None:
        writer = cv2.VideoWriter(
            "output_dlstreamer.mp4", cv2.VideoWriter_fourcc(*"mp4v"),
            30.0, (width, height))
    writer.write(frame)

pipeline.set_state(Gst.State.NULL)
if writer:
    writer.release()
print(f"Total text regions across all frames: {total_regions}", flush=True)

Device targets:

• "CPU" -- default for OpenVINO inference inside the appsink loop.
• "GPU" -- change "CPU" to "GPU" in core.compile_model().

Expected Output

---

License

Licensed under the MIT License. See LICENSE for details.

References

• PaddleOCR PP-OCRv4
• PaddleOCR OpenVINO Deployment
• OpenVINO Documentation
• Intel DLStreamer